Shape Optimizations Using AI Agents to Accelerate Numerical Flow Field Simulation and to Control the Island Model for Massive Parallelization

The overall aim of the project is to support the already automated design procedure for turbomachinery applications by extending individual steps within the workflow. Three main aspects are:

• Optimization of fluid mechanically relevant systems consisting of several components by respecting multiple objective criteria
• Reduction of the required computation time and overall number of individuals for the whole process by better initialization, supported mesh generation and improved parameterization
• Replacement of rule-based fitness functions by data-driven functions and their flexible combinations

A short overview about goals achieved during the first phase:

• Development of AI-supported agents for the automated generation of block structures in complex geometries.
• The implementation of an AI-based sensitivity analysis using back propagation within a neural network.
• AI-driven evaluation of flow fields integrated into axial turbine optimization.
• Implementation of a mixed optimizer strategy combining global evolutionary and local optimization, leveraging surrogate models to approximate gradients and assess convergence of the Pareto front (collaborative project with Prof. Peitz, Paderborn University).

• Assign different fitness functions to different islands – purposeful?
• Change fitness function in course of optimization – better overall results?

• Omit simple randomized search of geometry parameter space
• Generate highly diverse and at the same time valid population for starting optimization process

• Automatic block-structure as well as mesh element distribution with high quality
• Improve creation of prisms in boundary layer for hybrid meshes?

• Evaluate flow field directly on the basis of pressure and velocity fields to rank them without evaluating the integral values, e.g. torque and efficiency?
• Automatic setup of simulations possible with agent’s knowledge?

• Turbomachinery benchmark
• Meshing:
  • Prof. Langer / Dr. Lüddecke (Braunschweig/Göttingen)
  • Profs. Goubergrits / Knosalla (Berlin)

• Optimization
• Migration
• Fitness functions
• AI methods
• Data

1. Ebel, H.; van Delden, J.; Lüddecke, T.; Borse, A.; Gulakala, R.; Stoffel, M.; Yadav, M.; Stender, M.; Schindler, L.; de Payrebrune, K. M.; Raff, M.; Remy, C. D.; Röder, B.; Raj, R.; Rentschler, T.; Tismer, A.; Riedelbauch, S.; Eberhard, P., 2024, Data Publishing in Mechanics and Dynamics: Challenges, Guidelines, and Examples from Engineering Design. arXiv. https://doi.org/10.48550/ARXIV.2410.18358
   
   
2. Eyselein, S.; Tismer, A.; Raj, R.; Rentschler, T.; Riedelbauch, S.: AI-based clustering of numerical flow fields for accelerating the optimization of an axial turbine, Energies 2025, 18, 677. https://doi.org/10.3390/en18030677
   
   
3. Raj, R; Tismer, A.; Gaisser, L.; Riedelbauch, S.: A deep learning approach to calculate elementary effects of Morris sensitivity analysis. Proceedings in Applied Mathematics and Mechanics, e202400104, 2024. https://doi.org/10.1002/pamm.202400104
   
   
4. Rentschler, T.; Berkemeier, M. B.; Fraas, S.; Tismer, A.; Raj, R.; Peitz, S.; Riedelbauch, S.: Multi-criteria hydraulic turbine optimization using a genetic algorithm and trust-region postprocessing. Proceedings in Applied Mathematics and Mechanics, e202400126, 2024. https://doi.org/10.1002/pamm.202400126
   
   
5. Rentschler, T.; Raj, R.; Axial Turbine Database, GitHub Repository: https://github.com/ihs-ustutt/axial_turbine_database, 2024, https://doi.org/10.5281/zenodo.14014525
   
   
6. Rentschler, T.: Mesh Optimization for Computational Fluid Dynamics using Graph-Based Neural Networks, Master’s thesis, 2023, Institute of Fluid Mechanics and Hydraulic Machinery, Stuttgart.
   
   
7. Rentschler, T.: Wie Maschinen einfache Gitter für komplexe CFD-Simulationen lernen, 2024. https://rentschlertobias.github.io/lectures/Technologiefelder
   
   
8. Rentschler, T.; Tismer, A.; Riedelbauch, S.: Frame Field Prediction for Quadrilateral Domain Partition. Proceedings of the Canadian Society for Mechanical Engineering, 32nd Annual Conference of the CFD Society of Canada, Canadian Society of Rheology Symposium (CSME/CFDSC/CSR), 363, 2025.
   
   
9. Tismer, A.; Riedelbauch, S.: Current Development Status of a Design Framework for Hydraulic Machines. Proceedings of the Canadian Society for Mechanical Engineering, 32nd Annual Conference of the CFD Society of Canada, Canadian Society of Rheology Symposium (CSME/CFDSC/CSR), 266, 2025.